src/gallium/drivers/freedreno/a6xx/fd6_const.c

   1 /*
   2  * Copyright (C) 2016 Rob Clark <robclark@freedesktop.org>
   3  * Copyright © 2018 Google, Inc.
   4  *
   5  * Permission is hereby granted, free of charge, to any person obtaining a
   6  * copy of this software and associated documentation files (the "Software"),
   7  * to deal in the Software without restriction, including without limitation
   8  * the rights to use, copy, modify, merge, publish, distribute, sublicense,
   9  * and/or sell copies of the Software, and to permit persons to whom the
  10  * Software is furnished to do so, subject to the following conditions:
  11  *
  12  * The above copyright notice and this permission notice (including the next
  13  * paragraph) shall be included in all copies or substantial portions of the
  14  * Software.
  15  *
  16  * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
  17  * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
  18  * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
  19  * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
  20  * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
  21  * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
  22  * SOFTWARE.
  23  */
  24
  25 #include "fd6_const.h"
  26 #include "fd6_pack.h"
  27
  28 #include "ir3_const.h"
  29
  30 /* regid:          base const register
  31  * prsc or dwords: buffer containing constant values
  32  * sizedwords:     size of const value buffer
  33  */
  34 static void
  35 fd6_emit_const(struct fd_ringbuffer *ring, gl_shader_stage type,
  36                 uint32_t regid, uint32_t offset, uint32_t sizedwords,
  37                 const uint32_t *dwords, struct pipe_resource *prsc)
  38 {
  39         if (prsc) {
  40                 struct fd_bo *bo = fd_resource(prsc)->bo;
  41
  42                 if (fd6_geom_stage(type)) {
  43                         OUT_PKT(ring, CP_LOAD_STATE6_GEOM,
  44                                         CP_LOAD_STATE6_0(
  45                                                         .dst_off     = regid/4,
  46                                                         .state_type  = ST6_CONSTANTS,
  47                                                         .state_src   = SS6_INDIRECT,
  48                                                         .state_block = fd6_stage2shadersb(type),
  49                                                         .num_unit    = DIV_ROUND_UP(sizedwords, 4)
  50                                                 ),
  51                                         CP_LOAD_STATE6_EXT_SRC_ADDR(
  52                                                         .bo          = bo,
  53                                                         .bo_offset   = offset
  54                                                 )
  55                                 );
  56                 } else {
  57                         OUT_PKT(ring, CP_LOAD_STATE6_FRAG,
  58                                         CP_LOAD_STATE6_0(
  59                                                         .dst_off     = regid/4,
  60                                                         .state_type  = ST6_CONSTANTS,
  61                                                         .state_src   = SS6_INDIRECT,
  62                                                         .state_block = fd6_stage2shadersb(type),
  63                                                         .num_unit    = DIV_ROUND_UP(sizedwords, 4)
  64                                                 ),
  65                                         CP_LOAD_STATE6_EXT_SRC_ADDR(
  66                                                         .bo          = bo,
  67                                                         .bo_offset   = offset
  68                                                 )
  69                                 );
  70                 }
  71         } else {
  72                 /* NOTE we cheat a bit here, since we know mesa is aligning
  73                  * the size of the user buffer to 16 bytes.  And we want to
  74                  * cut cycles in a hot path.
  75                  */
  76                 uint32_t align_sz = align(sizedwords, 4);
  77                 dwords = (uint32_t *)&((uint8_t *)dwords)[offset];
  78
  79                 if (fd6_geom_stage(type)) {
  80                         OUT_PKTBUF(ring, CP_LOAD_STATE6_GEOM, dwords, align_sz,
  81                                         CP_LOAD_STATE6_0(
  82                                                         .dst_off     = regid/4,
  83                                                         .state_type  = ST6_CONSTANTS,
  84                                                         .state_src   = SS6_DIRECT,
  85                                                         .state_block = fd6_stage2shadersb(type),
  86                                                         .num_unit    = DIV_ROUND_UP(sizedwords, 4)
  87                                                 ),
  88                                         CP_LOAD_STATE6_1(),
  89                                         CP_LOAD_STATE6_2()
  90                                 );
  91                 } else {
  92                         OUT_PKTBUF(ring, CP_LOAD_STATE6_FRAG, dwords, align_sz,
  93                                         CP_LOAD_STATE6_0(
  94                                                         .dst_off     = regid/4,
  95                                                         .state_type  = ST6_CONSTANTS,
  96                                                         .state_src   = SS6_DIRECT,
  97                                                         .state_block = fd6_stage2shadersb(type),
  98                                                         .num_unit    = DIV_ROUND_UP(sizedwords, 4)
  99                                                 ),
 100                                         CP_LOAD_STATE6_1(),
 101                                         CP_LOAD_STATE6_2()
 102                                 );
 103                 }
 104         }
 105 }
 106
 107 static bool
 108 is_stateobj(struct fd_ringbuffer *ring)
 109 {
 110         return true;
 111 }
 112
 113 void
 114 emit_const(struct fd_ringbuffer *ring,
 115                 const struct ir3_shader_variant *v, uint32_t dst_offset,
 116                 uint32_t offset, uint32_t size, const void *user_buffer,
 117                 struct pipe_resource *buffer)
 118 {
 119         /* TODO inline this */
 120         assert(dst_offset + size <= v->constlen * 4);
 121         fd6_emit_const(ring, v->type, dst_offset,
 122                         offset, size, user_buffer, buffer);
 123 }
 124
 125 static void
 126 emit_const_bo(struct fd_ringbuffer *ring,
 127                 const struct ir3_shader_variant *v, uint32_t dst_offset,
 128                 uint32_t num, struct pipe_resource **prscs, uint32_t *offsets)
 129 {
 130         unreachable("shouldn't be called on a6xx");
 131 }
 132
 133 static void
 134 emit_tess_bos(struct fd_ringbuffer *ring, struct fd6_emit *emit, struct ir3_shader_variant *s)
 135 {
 136         struct fd_context *ctx = emit->ctx;
 137         const struct ir3_const_state *const_state = ir3_const_state(s);
 138         const unsigned regid = const_state->offsets.primitive_param * 4 + 4;
 139         uint32_t dwords = 16;
 140
 141         OUT_PKT7(ring, fd6_stage2opcode(s->type), 3);
 142         OUT_RING(ring, CP_LOAD_STATE6_0_DST_OFF(regid / 4) |
 143                         CP_LOAD_STATE6_0_STATE_TYPE(ST6_CONSTANTS)|
 144                         CP_LOAD_STATE6_0_STATE_SRC(SS6_INDIRECT) |
 145                         CP_LOAD_STATE6_0_STATE_BLOCK(fd6_stage2shadersb(s->type)) |
 146                         CP_LOAD_STATE6_0_NUM_UNIT(dwords / 4));
 147         OUT_RB(ring, ctx->batch->tess_addrs_constobj);
 148 }
 149
 150 static void
 151 emit_stage_tess_consts(struct fd_ringbuffer *ring, struct ir3_shader_variant *v,
 152                 uint32_t *params, int num_params)
 153 {
 154         const struct ir3_const_state *const_state = ir3_const_state(v);
 155         const unsigned regid = const_state->offsets.primitive_param;
 156         int size = MIN2(1 + regid, v->constlen) - regid;
 157         if (size > 0)
 158                 fd6_emit_const(ring, v->type, regid * 4, 0, num_params, params, NULL);
 159 }
 160
 161 static void
 162 emit_tess_consts(struct fd6_emit *emit)
 163 {
 164         struct fd_context *ctx = emit->ctx;
 165
 166         struct fd_ringbuffer *constobj = fd_submit_new_ringbuffer(
 167                 ctx->batch->submit, 0x1000, FD_RINGBUFFER_STREAMING);
 168
 169         /* VS sizes are in bytes since that's what STLW/LDLW use, while the HS
 170          * size is dwords, since that's what LDG/STG use.
 171          */
 172         unsigned num_vertices =
 173                 emit->hs ?
 174                 emit->info->vertices_per_patch :
 175                 emit->gs->shader->nir->info.gs.vertices_in;
 176
 177         uint32_t vs_params[4] = {
 178                 emit->vs->shader->output_size * num_vertices * 4,       /* vs primitive stride */
 179                 emit->vs->shader->output_size * 4,                                      /* vs vertex stride */
 180                 0,
 181                 0
 182         };
 183
 184         emit_stage_tess_consts(constobj, emit->vs, vs_params, ARRAY_SIZE(vs_params));
 185
 186         if (emit->hs) {
 187                 uint32_t hs_params[4] = {
 188                         emit->vs->shader->output_size * num_vertices * 4,       /* vs primitive stride */
 189                         emit->vs->shader->output_size * 4,                                      /* vs vertex stride */
 190                         emit->hs->shader->output_size,
 191                         emit->info->vertices_per_patch
 192                 };
 193
 194                 emit_stage_tess_consts(constobj, emit->hs, hs_params, ARRAY_SIZE(hs_params));
 195                 emit_tess_bos(constobj, emit, emit->hs);
 196
 197                 if (emit->gs)
 198                         num_vertices = emit->gs->shader->nir->info.gs.vertices_in;
 199
 200                 uint32_t ds_params[4] = {
 201                         emit->ds->shader->output_size * num_vertices * 4,       /* ds primitive stride */
 202                         emit->ds->shader->output_size * 4,                                      /* ds vertex stride */
 203                         emit->hs->shader->output_size,                      /* hs vertex stride (dwords) */
 204                         emit->hs->shader->nir->info.tess.tcs_vertices_out
 205                 };
 206
 207                 emit_stage_tess_consts(constobj, emit->ds, ds_params, ARRAY_SIZE(ds_params));
 208                 emit_tess_bos(constobj, emit, emit->ds);
 209         }
 210
 211         if (emit->gs) {
 212                 struct ir3_shader_variant *prev;
 213                 if (emit->ds)
 214                         prev = emit->ds;
 215                 else
 216                         prev = emit->vs;
 217
 218                 uint32_t gs_params[4] = {
 219                         prev->shader->output_size * num_vertices * 4,   /* ds primitive stride */
 220                         prev->shader->output_size * 4,                                  /* ds vertex stride */
 221                         0,
 222                         0,
 223                 };
 224
 225                 num_vertices = emit->gs->shader->nir->info.gs.vertices_in;
 226                 emit_stage_tess_consts(constobj, emit->gs, gs_params, ARRAY_SIZE(gs_params));
 227         }
 228
 229         fd6_emit_take_group(emit, constobj, FD6_GROUP_PRIMITIVE_PARAMS, ENABLE_ALL);
 230 }
 231
 232 static void
 233 fd6_emit_ubos(struct fd_context *ctx, const struct ir3_shader_variant *v,
 234                 struct fd_ringbuffer *ring, struct fd_constbuf_stateobj *constbuf)
 235 {
 236         if (!v->shader->num_ubos)
 237                 return;
 238
 239         int num_ubos = v->shader->num_ubos;
 240
 241         OUT_PKT7(ring, fd6_stage2opcode(v->type), 3 + (2 * num_ubos));
 242         OUT_RING(ring, CP_LOAD_STATE6_0_DST_OFF(0) |
 243                         CP_LOAD_STATE6_0_STATE_TYPE(ST6_UBO)|
 244                         CP_LOAD_STATE6_0_STATE_SRC(SS6_DIRECT) |
 245                         CP_LOAD_STATE6_0_STATE_BLOCK(fd6_stage2shadersb(v->type)) |
 246                         CP_LOAD_STATE6_0_NUM_UNIT(num_ubos));
 247         OUT_RING(ring, CP_LOAD_STATE6_1_EXT_SRC_ADDR(0));
 248         OUT_RING(ring, CP_LOAD_STATE6_2_EXT_SRC_ADDR_HI(0));
 249
 250         for (int i = 0; i < num_ubos; i++) {
 251                 struct pipe_constant_buffer *cb = &constbuf->cb[i];
 252
 253                 /* If we have user pointers (constbuf 0, aka GL uniforms), upload them
 254                  * to a buffer now, and save it in the constbuf so that we don't have
 255                  * to reupload until they get changed.
 256                  */
 257                 if (cb->user_buffer) {
 258                         struct pipe_context *pctx = &ctx->base;
 259                         u_upload_data(pctx->stream_uploader, 0,
 260                                         cb->buffer_size,
 261                                         64,
 262                                         cb->user_buffer,
 263                                         &cb->buffer_offset, &cb->buffer);
 264                         cb->user_buffer = NULL;
 265                 }
 266
 267                 if (cb->buffer) {
 268                         int size_vec4s = DIV_ROUND_UP(cb->buffer_size, 16);
 269                         OUT_RELOC(ring, fd_resource(cb->buffer)->bo,
 270                                         cb->buffer_offset,
 271                                         (uint64_t)A6XX_UBO_1_SIZE(size_vec4s) << 32,
 272                                         0);
 273                 } else {
 274                         OUT_RING(ring, 0xbad00000 | (i << 16));
 275                         OUT_RING(ring, 0xbad00000 | (i << 16));
 276                 }
 277         }
 278 }
 279
 280 static unsigned
 281 user_consts_cmdstream_size(struct ir3_shader_variant *v)
 282 {
 283         struct ir3_ubo_analysis_state *ubo_state = &ir3_const_state(v)->ubo_state;
 284
 285         if (unlikely(!ubo_state->cmdstream_size)) {
 286                 unsigned packets, size;
 287
 288                 /* pre-calculate size required for userconst stateobj: */
 289                 ir3_user_consts_size(ubo_state, &packets, &size);
 290
 291                 /* also account for UBO addresses: */
 292                 packets += 1;
 293                 size += 2 * v->shader->num_ubos;
 294
 295                 unsigned sizedwords = (4 * packets) + size;
 296                 ubo_state->cmdstream_size = sizedwords * 4;
 297         }
 298
 299         return ubo_state->cmdstream_size;
 300 }
 301
 302 static void
 303 emit_user_consts(struct fd6_emit *emit)
 304 {
 305         static const enum pipe_shader_type types[] = {
 306                         PIPE_SHADER_VERTEX, PIPE_SHADER_TESS_CTRL, PIPE_SHADER_TESS_EVAL,
 307                         PIPE_SHADER_GEOMETRY, PIPE_SHADER_FRAGMENT,
 308         };
 309         struct ir3_shader_variant *variants[] = {
 310                         emit->vs, emit->hs, emit->ds, emit->gs, emit->fs,
 311         };
 312         struct fd_context *ctx = emit->ctx;
 313         unsigned sz = 0;
 314
 315         for (unsigned i = 0; i < ARRAY_SIZE(types); i++) {
 316                 if (!variants[i])
 317                         continue;
 318                 sz += user_consts_cmdstream_size(variants[i]);
 319         }
 320
 321         struct fd_ringbuffer *constobj = fd_submit_new_ringbuffer(
 322                         ctx->batch->submit, sz, FD_RINGBUFFER_STREAMING);
 323
 324         for (unsigned i = 0; i < ARRAY_SIZE(types); i++) {
 325                 if (!variants[i])
 326                         continue;
 327                 ir3_emit_user_consts(ctx->screen, variants[i], constobj, &ctx->constbuf[types[i]]);
 328                 fd6_emit_ubos(ctx, variants[i], constobj, &ctx->constbuf[types[i]]);
 329         }
 330
 331         fd6_emit_take_group(emit, constobj, FD6_GROUP_CONST, ENABLE_ALL);
 332 }
 333
 334 void
 335 fd6_emit_consts(struct fd6_emit *emit)
 336 {
 337         struct fd_context *ctx = emit->ctx;
 338         struct fd6_context *fd6_ctx = fd6_context(ctx);
 339
 340         if (emit->dirty & (FD_DIRTY_CONST | FD_DIRTY_PROG))
 341                 emit_user_consts(emit);
 342
 343         if (emit->key.key.has_gs || emit->key.key.tessellation)
 344                 emit_tess_consts(emit);
 345
 346         /* if driver-params are needed, emit each time: */
 347         const struct ir3_shader_variant *vs = emit->vs;
 348         if (ir3_needs_vs_driver_params(vs)) {
 349                 struct fd_ringbuffer *dpconstobj = fd_submit_new_ringbuffer(
 350                                 ctx->batch->submit, IR3_DP_VS_COUNT * 4, FD_RINGBUFFER_STREAMING);
 351                 ir3_emit_vs_driver_params(vs, dpconstobj, ctx, emit->info);
 352                 fd6_emit_take_group(emit, dpconstobj, FD6_GROUP_VS_DRIVER_PARAMS, ENABLE_ALL);
 353                 fd6_ctx->has_dp_state = true;
 354         } else if (fd6_ctx->has_dp_state) {
 355                 fd6_emit_take_group(emit, NULL, FD6_GROUP_VS_DRIVER_PARAMS, ENABLE_ALL);
 356                 fd6_ctx->has_dp_state = false;
 357         }
 358 }
 359
 360 void
 361 fd6_emit_ibo_consts(struct fd6_emit *emit, const struct ir3_shader_variant *v,
 362                 enum pipe_shader_type stage, struct fd_ringbuffer *ring)
 363 {
 364         struct fd_context *ctx = emit->ctx;
 365
 366         ir3_emit_ssbo_sizes(ctx->screen, v, ring, &ctx->shaderbuf[stage]);
 367         ir3_emit_image_dims(ctx->screen, v, ring, &ctx->shaderimg[stage]);
 368 }
 369
 370 void
 371 fd6_emit_cs_consts(const struct ir3_shader_variant *v, struct fd_ringbuffer *ring,
 372                 struct fd_context *ctx, const struct pipe_grid_info *info)
 373 {
 374         ir3_emit_cs_consts(v, ring, ctx, info);
 375         fd6_emit_ubos(ctx, v, ring, &ctx->constbuf[PIPE_SHADER_COMPUTE]);
 376 }
 377
 378 void
 379 fd6_emit_immediates(struct fd_screen *screen, const struct ir3_shader_variant *v,
 380                 struct fd_ringbuffer *ring)
 381 {
 382         ir3_emit_immediates(screen, v, ring);
 383 }
 384
 385 void
 386 fd6_emit_link_map(struct fd_screen *screen,
 387                 const struct ir3_shader_variant *producer,
 388                 const struct ir3_shader_variant *v, struct fd_ringbuffer *ring)
 389 {
 390         ir3_emit_link_map(screen, producer, v, ring);
 391 }